Skip to main content
A conversation is one end-to-end interaction — a phone call, a chat thread, a voicemail. Each conversation belongs to one agent and owns:
  • A transcripts row with the full turn-by-turn content.
  • An optional recording (audio for voice, transcript-only for chat).
  • Zero or more conversation_analyses — the assessment saved each time the conversation is analyzed.
  • Zero or more analysis_findings — per-conversation observations, each tagged with a finding type.
Every conversation also carries the basics you see in the list: the agent, its company attribution when present, a reference id (the provider’s conversation id), the ended reason, duration, cost, and whether it arrived from a provider webhook or a direct upload.

Summaries

The conversation detail page can generate a short AI summary on demand. Click Generate summary and Zelto summarizes the transcript and caches the result on the row; opening the conversation again reads the cached summary instead of regenerating it.

Monitor results

The conversation detail page lists every monitor assigned to the conversation’s agent. For each monitor, it shows the conversation’s affected status, score, stage outputs, and reasoning when the evaluation is available. A monitor shows N/A when it is assigned to the agent but has not evaluated that conversation yet. If the agent has no assigned monitors, the panel also shows N/A.

Tool calls

When an agent invokes a tool — a function call such as checking availability, booking an appointment, or looking up an order — Zelto captures each call and puts it to work three ways:
  • Inline in the transcript. The call shows up as a tool turn in the conversation timeline, in order, with the tool name, its arguments, the result, and a status badge (success / error / pending). Open a conversation to read tool calls alongside what was said.
  • In a tool-check monitor. A tool_check stage classifies the conversation from those events — whether an expected tool fired, succeeded, errored, returned empty, or fired twice. Confirmation language in the transcript is a claim, never treated as successful execution.
  • In the AI analysis. Summaries, findings, scorecards, and metrics also read the structured call — name, arguments, and result — so a model stage can reason about what the agent actually did (e.g. “booked the wrong date”), not just what it said. Use a tool check, not an AI stage, when the question is whether a specific tool executed.
  • In tool-call analytics. Every tool call is also written to a tool_call_traces row (tool name, arguments, result, status, timing) so tool usage, error rates, and latency can be analyzed across many conversations — not just read one conversation at a time. Open Tool calls to see that organization-wide view.
Tool calls are captured automatically for the providers that emit them — Vapi, Retell, ElevenLabs, and LiveKit. If you push conversations yourself, attach a toolCall to any tool turn — see Custom & other providers. Capture is forward-looking: conversations ingested from when the feature ships onward include their tool calls.

Filtering

The conversations list filters on: Filters persist in the URL, so a filtered list is shareable. Metadata filters are property rows in the style of a product-analytics tool: a key (a dotted path into the conversation’s provider metadata, such as campaign, model.provider, or for Retell calls retell.transfer and retell.call_analysis.custom_analysis_data.<field>; keys seen on your recent conversations are suggested), an operator (equals, does not equal, contains, does not contain, numeric greater than / less than and their or equal forms, is set, is not set, is true, is false), and a value. Up to eight rows combine with AND, and the set is encoded in the URL as ?meta=key:op:value. The same bar scopes the numbers on the Monitors list, Compare view, and a monitor’s Overview and Results tabs. Field suggestions are discovered automatically from your metadata, including nested paths such as contexto.call_id; no customer-specific setup is required. To find a call by that property, choose contexto.call_id → equals → the call ID. Numeric and boolean values can also be matched by their text representation. Missing properties keep the existing is set / is not set behavior. Selective exact-value filters (eq / neq) use the metadata search index once historical coverage is ready. Common values and other operators retain the JSON query path, avoiding large historical candidate sets. contains and does not contain keep their existing case-insensitive substring behavior; narrow large histories with a date range for those searches. REST and MCP accept the same filter, for example meta=contexto.call_id:eq:1234567890. Selecting a company in the sidebar applies the same exact per-conversation filter. This is more precise than agent-based company views when one agent serves several companies.

Findings

Any conversation can be linked to a finding — the durable, shareable issue it’s an example of — and curated findings can be pushed to Linear from the finding’s panel. The assessment saved when a conversation is analyzed lives in its conversation_analyses row. For the OpenTelemetry view of a conversation — the span tree, per-step latency, tokens, and logs — see Traces. A trace links to its conversation when their session ids match.

Who read this conversation

Opening a conversation, and fetching its transcript, writes a row to your org’s audit log — the actor, the timestamp, and the source IP. Owners and admins can read that trail under Settings → Audit log, which also lists the times Zelto staff opened one of your conversations.